The MAD Podcast with Matt Turck

Retrieval-Augmented Generation (RAG) has become a dominant architecture in modern AI deployments, and in this episode, we sit down with Douwe Kiela, who co-authored the original RAG paper in 2020. Douwe is now the founder and CEO of Contextual AI, a startup focusing on helping enterprises deploy RAG as an agentic system. We start the conversation with Douwe's thoughts on the very latest advancements in Generative AI, including GPT 4.5, DeepSeek and the exciting paradigm shift towards test time compute, as well as the US-China rivalry in AI. We then dive into RAG: definition, origin story and core architecture. Douwe explains the evolution of RAG into RAG 2.0 and Agentic RAG, emphasizing the importance of self-learning systems over individual models and the role of synthetic data. We close with the challenges and opportunities of deploying AI in real-world enterprise, discussing the balance between accuracy and the inherent inaccuracies of AI systems.Contextual AIWebsite - https://contextual.aiX/Twitter - https://x.com/ContextualAIDouwe KielaLinkedIn - https://www.linkedin.com/in/douwekielaX/Twitter - https://x.com/douwekielaFIRSTMARKWebsite - https://firstmark.comX/Twitter - https://twitter.com/FirstMarkCapMatt Turck (Managing Director)LinkedIn - https://www.linkedin.com/in/turck/X/Twitter - https://twitter.com/mattturck(00:00) Intro(01:57) Thoughts on the latest AI models: GPT-4.5, Sonnet 3.7, Grok 3(04:50) The test time compute paradigm shift(06:47) Unsupervised learning vs reasoning: a false dichotomy(07:30) The significance of DeepSeek(10:29) USA vs. China: is the AI war overblown?(12:19) Controlling AI hallucinations at the model level(13:51) RAG: definition and origin story(18:46) Why the Transformers paper initially felt underwhelming(20:41) The core architecture of RAG(26:06) RAG vs. fine-tuning vs. long context windows(30:53) RAG 2.0: Thinking in systems and not models(31:28) Data extraction and data curation for RAG(35:59) Contextual Language Models (CLMs)(38:04) Finetuning and alignment techniques: GRIT, KTO, LENS(40:40) Agentic RAG(41:36) General vs. specialized RAG agents(44:35) Synthetic data in AI(45:51) Deploying AI in the enterprise(48:07) How tolerant are enterprises to AI hallucinations?(49:35) The future of Contextual AI

In this episode, we dive into how AI is transforming video editing with Gaurav Misra, the CEO of Captions. Launched in New York in 2021, Captions already empowers over 10 million creators worldwide, leveraging AI to make video production as simple as clicking a button.Discover the strategic framework that led to the inception of Captions, and learn how the founders identified societal changes and technological advancements to build a groundbreaking company. We explore the challenges and opportunities of building an AI product for video editing, including how Captions is outpacing traditional content production workflows.Gaurav shares insights into the future of video editing, the role of AI in democratizing video production, and the unique approach Captions takes to differentiate itself from industry giants like Adobe and Capcut. CaptionsWebsite - https://www.captions.aiX/Twitter - https://x.com/getcaptionsappGaurav MisraLinkedIn - https://www.linkedin.com/in/gamisra1X/Twitter - https://x.com/gmharharFIRSTMARKWebsite - https://firstmark.comX/Twitter - https://twitter.com/FirstMarkCapMatt Turck (Managing Director)LinkedIn - https://www.linkedin.com/in/turck/X/Twitter - https://twitter.com/mattturck(00:00) Intro(01:30) What is Captions?(03:43) How did Captions start?(08:25) The strategy behind launching Captions(12:32) How is Captions different from other editing tools?(14:13) How does it compare to CapCut?(18:22) Who is the typical Captions user?(20:13) Why ‘Captions’?(23:47) Captions’ product suite for production and editing(26:37) AI models powering Captions(36:22) AI lipsync(38:49) Personalized fine-tuned models for creators?(39:38) Building models vs. building wrappers(43:09) Cloud AI vs. Local AI(45:19) Optimizing for low latency(48:07) AI/ML stack at Captions(51:10) “Hallucinations are a feature, not a bug”(53:19) Prompt engineering(54:12) Have we passed the uncanny valley for AI avatars?(01:01:47) The impact of deepfakes(01:04:33) CapCut ban and its effects(01:05:05) Evolving from paid to freemium(01:07:42) Building a company on foundation models(01:09:01) Running an AI company in New York

AI customer service agents are quickly replacing the often clunky AI chatbots of years past, and revolutionizing how we all interact with customer service. In this episode, we dive into this rapid transformation with Mike Murchison, CEO of Ada, a fast-growing leader in the space.Mike shares how harnessing the power of several Generative AI models enables Ada to automate up to 83% of customer interactions, providing a seamless and empathetic service that rivals, and will soon surpass, human agents. We explore the challenges and triumphs of deploying AI in customer service in this new era, from the intricacies of model orchestration to the importance of resolution and empathy. Mike also teases the future of agentic AI in the enterprise, where AI agents collaborate across departments to innovate and improve products.AdaWebsite - https://www.ada.cxX/Twitter - https://x.com/ada_cxMike MurchisonLinkedIn - https://www.linkedin.com/in/mikemurchisonX/Twitter - https://x.com/mimurchisonFIRSTMARKWebsite - https://firstmark.comX/Twitter - https://twitter.com/FirstMarkCapMatt Turck (Managing Director)LinkedIn - https://www.linkedin.com/in/turck/X/Twitter - https://twitter.com/mattturck(00:00) Intro(02:27) Why is customer service a perfect use case for AI?(03:36) Why didn’t foundation models replace AI “thin wrappers” out of the box?(05:27) What is Ada?(10:41) Reasoning engine, model orchestration, instruction following, routing(15:45) Hybrid systems, finetuning, customization(18:28) Prompt engineering, observability, self-improvement(22:07) RAG (Retrieval-Augmented Generation) and AI as a judge(23:06) Guardrails and security(24:33) Should we expect perfection from AI?(26:14) Measuring “resolution”(29:29) What actions can Ada AI Agents take?(32:12) Authentication and personalization(35:09) Handoff vs human delegation(38:12) ACX (AI Customer Experience) and the future of customer service professionals(42:13) Leveraging analytics and customer support data(45:54) AI agents for cross-selling and upselling(48:25) Traditional AI chatbots vs the new generation of AI Agents(51:24) Emotion, empathy, personality(54:56) Transparency and AI improvement(57:58) Managing AI: the measure-coach-improve loop(1:00:15) Ada Voice and Email(1:06:25) Future predictions for AI(1:07:56) Multi-agent collaboration

Replit is one of the most visible and exciting companies reshaping how we approach software and application development in the Generative AI era. In this episode, we sit down with its CEO, Amjad Masad, for an in-depth discussion on all things AI, agents, and software. Amjad shares the journey of building Replit, from its humble beginnings as a student side project to becoming a major player in Generative AI today. We also discuss the challenges of launching a startup, the multiple attempts to get into Y Combinator, the pivotal moment when Paul Graham recognized Replit’s potential, and the early bet on integrating AI and machine learning into the core of Replit. Amjad dives into the evolving landscape of AI and machine learning, sharing how these technologies are reshaping software development. We explore the concept of coding agents and the impact of Replit’s latest innovation, Replit Agent, on the software creation process. Additionally, Amjad reflects on his time at Codecademy and Facebook, where he worked on groundbreaking projects like React Native, and how those experiences shaped his entrepreneurial journey. We end with Amjad's view on techno-optimism and his belief in an energized Silicon Valley. Replit Website - https://replit.com X/Twitter - https://x.com/Replit Amjad Masad LinkedIn - https://www.linkedin.com/in/amjadmasad X/Twitter - https://x.com/amasad FIRSTMARK Website - https://firstmark.com X/Twitter - https://twitter.com/FirstMarkCap Matt Turck (Managing Director) LinkedIn - https://www.linkedin.com/in/turck/ X/Twitter - https://twitter.com/mattturck (00:00) Intro (01:36) The origins of Replit (15:54) Amjad’s decision to restart Replit (19:00) Joining Y Combinator (30:06) AI and ML at Replit (32:31) Explain Code (39:09) Replit Agent (52:10) Balancing usability for both developers and non-technical users (53:22) Sonnet 3.5 stack (58:43) The challenge of AI evaluation (01:00:02) ACI vs. HCI (01:05:02) Will AI replace software development? (01:10:15) If anyone can build an app with Replit, what’s the next bottleneck? (01:14:31) The future of SaaS in an AI-driven world (01:18:37) Why Amjad embraces techno-optimism (01:20:36) Defining civilizationism (01:23:11) Amjad’s perspective on government’s role

In this episode, we explore the cutting-edge world of data infrastructure with Justin Borgman, CEO of Starburst — a company transforming data analytics through its open-source project, Trino, and empowering industry giants like Netflix, Airbnb, and LinkedIn. Justin takes us through Starburst’s journey from a Yale University spin-out to a leading force in data innovation, discussing the shift from data lakes to lakehouses, the rise of open formats like Iceberg as the future of data storage, and the role of AI in modern data applications. We also dive into how Starburst is staying ahead by balancing on-prem and cloud offerings while emphasizing the value of optionality in a rapidly evolving, data-driven landscape. Starburst Data Website - https://www.starburst.io X/Twitter - https://x.com/starburstdata Justin Borgman LinkedIn - https://www.linkedin.com/in/justinborgman X/Twitter - https://x.com/justinborgman FIRSTMARK Website - https://firstmark.com X/Twitter - https://twitter.com/FirstMarkCap Matt Turck (Managing Director) LinkedIn - https://www.linkedin.com/in/turck/ X/Twitter - https://twitter.com/mattturck (00:00) Intro (01:32) What is Starburst? (02:32) Understanding the data layer (05:06) Justin Borgman’s story before Starburst (10:41) The evolution of Presto into Trino (13:20) Lakehouse vs. data lake vs. data warehouse (22:06) Why Starburst backed the lakehouse from the start (23:20) Starburst Enterprise (27:31) Cloud vs. on-prem (29:10) Starburst Galaxy (31:23) Dell Data Lakehouse (32:13) Starburst’s data architecture explained (38:30) The rise of data apps (38:54) Starburst AML (40:41) “We actually built the Galaxy twice” (43:13) Managing multiple products at scale (45:14) “We founded the company on the idea of optionality” (47:20) Iceberg (48:01) How open-source acquisitions work (51:39) Why Snowflake embraced Iceberg (53:15) Data mesh (55:31) AI at Starburst (57:16) Key takeaways from go-to-market strategies (01:01:18) Lessons from the Dell partnership (01:04:40) Predictions for 2025

As AI takes over the world, data is more than ever “the new oil”, and data engineering is the discipline that makes data usable behind the scenes. In this episode, we dive deep into the present and future of data engineering with Ben Rogojan, also known as the Seattle Data Guy. A seasoned data engineering consultant, Ben has built a big brand and reputation in the field with over 100k followers on platforms like YouTube and Substack. We started the conversation with a deep dive into data engineering as a profession: what do data engineers actually do? What is the career path, and what should aspiring data engineers learn? We then explored some of the biggest stories of 2024 (including the rise of Iceberg) and went into some predictions for 2025, as a way to discuss some key topics everyone should be familiar with in data engineering, including the integration of AI in data workflows, the potential for automation, and why SQL isn't going anywhere. Discover how companies are navigating the complexities of data infrastructure, the rise of open table formats like Iceberg, and the ongoing battle between data giants like Snowflake and Databricks. Ben Rogojan Website - https://www.theseattledataguy.com Newsletter - https://seattledataguy.substack.com LinkedIn - https://www.linkedin.com/company/seattle-data-guy X/Twitter - https://x.com/seattledataguy FIRSTMARK Website - https://firstmark.com X/Twitter - https://twitter.com/FirstMarkCap Matt Turck (Managing Director) LinkedIn - https://www.linkedin.com/in/turck/ X/Twitter - https://twitter.com/mattturck (00:00) Intro (01:20) Why 2025 will be huge for data engineering (02:55) The story of the Seattle Data Guy (06:51) What exactly is data engineering? (07:41) Data, AI, and ML: where do they overlap? (09:23) Data analyst vs. data engineer vs. data scientist: what’s the difference? (11:20) A day in the life of a data engineer (12:58) Data engineering: Silicon Valley vs. everywhere else (15:27) How to become an AI engineer (28:46) Will AI replace AI engineers? (33:42) Why is the data world so complex? (36:53) The functional consolidation of the data world (38:34) Big data stories from 2024 (39:28) Why Iceberg is a game-changer (46:02) How startups manage data in their early days (48:44) Seattle Data Guy’s favorite tools (50:09) Bold predictions for 2025

In this episode, we dive deep into the world of AI engineering with Chip Huyen, author of the excellent, newly released book "AI Engineering: Building Applications with Foundation Models". We explore the nuances of AI engineering, distinguishing it from traditional machine learning, discuss how foundational models make it possible for anyone to build AI applications and cover many other topics including the challenges of AI evaluation, the intricacies of the generative AI stack, why prompt engineering is underrated, why the rumors of the death of RAG are greatly exaggerated, and the latest progress in AI agents. Book: https://www.oreilly.com/library/view/ai-engineering/9781098166298/ Chip Huyen Website - https://huyenchip.com LinkedIn - https://www.linkedin.com/in/chiphuyen Twitter/X - https://x.com/chipro FIRSTMARK Website - https://firstmark.com Twitter - https://twitter.com/FirstMarkCap Matt Turck (Managing Director) LinkedIn - https://www.linkedin.com/in/turck/ Twitter - https://twitter.com/mattturck (00:00) Intro (02:45) What is new about AI engineering? (06:11) The product-first approach to building AI applications (07:38) Are AI engineering and ML engineering two separate professions? (11:00) The Generative AI stack (13:00) Why are language models able to scale? (14:45) Auto-regressive vs. masked models (16:46) Supervised vs. unsupervised vs. self-supervised (18:56) Why does model scale matter? (20:40) Mixture of Experts (24:20) Pre-training vs. post-training (28:43) Sampling (32:14) Evaluation as a key to AI adoption (36:03) Entropy (40:05) Evaluating AI systems (43:21) AI as a judge (46:49) Why prompt engineering is underrated (49:38) In-context learning (51:46) Few-shot learning and zero-shot learning (52:57) Defensive prompt engineering (55:29) User prompt vs. system prompt (57:07) Why RAG is here to stay (01:00:31) Defining AI agents (01:04:04) AI agent planning (01:08:32) Training data as a bottleneck to agent planning

In this episode, we sit down with Florian Douetteau, co-founder and CEO of Dataiku, a global category leader in enterprise AI and a fixture on the Forbes Cloud 100 list and in the Gartner Leader Quadrant. Florian shares his journey from a Parisian student fascinated by functional programming to leading a global enterprise software company. We discuss how Dataiku bridges the gap between technical and business teams to democratize AI in the enterprise, the challenges of selling to enterprise clients, and how Dataiku acts as an orchestration layer for Generative AI, helping businesses manage complex data processes and control AI, so they can build more with AI. Dataiku Website - https://www.dataiku.com/ X/Twitter - https://twitter.com/dataiku Florian Douetteau LinkedIn - https://www.linkedin.com/in/fdouetteau X/Twitter - https://twitter.com/fdouetteau FIRSTMARK Website - https://firstmark.com X/Twitter - https://twitter.com/FirstMarkCap Matt Turck (Managing Director) LinkedIn - https://www.linkedin.com/in/turck/ X/Twitter - https://twitter.com/mattturck (00:00) Intro (02:08) Florian's life before Dataiku (06:58) Creation of Dataiku (12:08) Secret behind the Dataiku's name (12:47) How does Dataiku stay insightful about the future? (14:46) Building a platform, not just a tool (17:26) How to sell to the enterprise from the beginning (20:09) Dataiku platform today (26:55) Data is always the problem (28:50) LLM Mesh (36:02) Will Gen AI replace ML? (39:41) Managing Gen AI and traditional AI on one platform (40:37) Gen AI deployment in the enterprise (48:33) Dataiku's roadmap (50:28) What has changed with the company's growth?

In this episode, we dive into the world of generative AI with May Habib, co-founder of Writer, a platform transforming enterprise AI use. May shares her journey from Qordoba to Writer, emphasizing the impact of transformers in AI. We explore Writer's graph-based RAG approach, and their AI Studio for building custom applications. We also discuss Writer's Autonomous Action functionality, set to revolutionize AI workflows by enabling systems to act autonomously, highlighting AI's potential to accelerate product development and market entry with significant increases in capacity and capability. Writer Website - https://writer.com X/Twitter - https://x.com/get_writer  May Habib LinkedIn - https://www.linkedin.com/in/may-habib X/Twitter - https://x.com/may_habib FIRSTMARK Website - https://firstmark.com X/Twitter - https://twitter.com/FirstMarkCap Matt Turck (Managing Director) LinkedIn - https://www.linkedin.com/in/turck/ X/Twitter - https://twitter.com/mattturck This session was recorded live at a recent Data Driven NYC, our in-person, monthly event series, hosted at Ramp's beautiful HQ. If you are ever in New York, you can join the upcoming events here: https://www.eventbrite.com/o/firstmark-capital-2215570183 (00:00) Intro (01:47) What is Writer? (02:52) Writer's founding story (06:54) Writer is a full-stack company. Why? (07:57) Writer's enterprise use cases (10:51) Knowledge Graph (17:59) Guardrails (20:17) AI Studio (23:16) Palmyra X 004 (27:18) Current state of the AI adoption in enterprises (28:57) Writer's sales approach (31:25) What May Habib is excited about in AI (33:14) Autonomous Action use cases

Nathan Benaich, founder and GP at VC firm Air Street Capital, publishes every year "State of AI", one of the most widely-read and comprehensive reports on all things AI across research, industry, and policy. In this episode, we sit down with Nathan to discuss some of the highlights of the 2024 edition of the report, including the "vibes" shift in the industry from existential risk concerns last year to the current monetization race, the financial success of the foundation model labs, how a generative AI app could top the Apple Store charts in 2025, and the challenges facing humanoid robotics. State of AI 2024 report: https://www.stateof.ai/2024-report-launch State of AI 2024 video: https://youtu.be/EVMbnPOuUl0 Air Street Capital Website - https://www.airstreet.com X/Twitter - https://x.com/airstreet Nathan Benaich LinkedIn - https://www.linkedin.com/in/nathanbenaich X/Twitter - https://x.com/nathanbenaich FirstMark Website - https://firstmark.com X/Twitter - https://twitter.com/FirstMarkCap Matt Turck (Managing Director) LinkedIn - https://www.linkedin.com/in/turck/ X/Twitter - https://twitter.com/mattturck (01:08) Who is Nathan Benaich? (04:57) "Vibe" shift in AI (09:13) Current state of the foundation models (22:01) AI companies vs. SaaS (23:31) AI consumer apps (25:49) AI applications from a VC's perspective (29:25) "You don't need to be an AI engineer to build an AI company" (30:46) AI in robotics (34:36) AI regulations in Europe (40:55) Predictions on the future of AI (49:30) Nathan Benaich's favorite sources of information